Hi.
I estimated a model in 2 different samples (each one in a different mod file). In each mod file I estimated the same 2 parameters and for the others parameters I used a set of values for calibration (the calibration is the same in the mod files).
The estimated values (and distributions - I’m using MCMC) for the 2 parameters are different between the samples as I expected, but the IRFs are EXACTLY the same.
When I compare the structure of IRFs with isequaln(oo1_.irfs,oo2_.irfs)
, I received a TRUE.
I have a stoch_simul
command after the estimation
command in both mod files.
I’ve also checked that oo_.dr.ghx and oo_.dr.ghu are the same in the two samples. The smoothed variables and smoothed shocks are different as expected.
Is this thing common/possible? I should expect that the IRFs are really similar, but there is no difference at all. I wonder if I miss something about Dynare.
Actually, this happened with 6 different versions of the model (3 versions are estimating 3 parameters and 3 versions , 2 parameters).
When I tried to estimate only 1 parameter, then the IRF are different (they are similar, but different, that’s what I was expecting).
Would you please help me with some advice?
Thanks in advance.
Is the final M_.params
different in the two results files? It may be that oo_.irfs
was not properly updated because the stoch_simul
-command did not work as expected.
Thanks for you answer, professor Pfeifer.
The values in M_.params are equal in two results files.
Actually, I’m little confused what’s stored in this vector. According to 4. The model file — Dynare 4.6.3 documentation, I understood that M_.params are the values of calibration. Despite this, in some other tests of mine, I’ve seen that the M_.params are updated if I use a estimated_params
block with priors. In that case M_.params are updated to the modes of priors.
If I understood you correctly, the M_.params should be updated again after a estimation command. What should be in M_.params in that case? The optimization value or MH mean (or mode)?
And what can I do to make stoch_simul work as expected?
Here are the files if you want to see them.
Files.zip (18.4 KB)
Thank you again.
Update: The “problem” is in identification part. When dynare enters in the identification_analysis
,
it updates the M_.params to the mode of the priors again, overwriting the posterior_mean. It occurs in the set_all_parameters
part (line 114 of identification_analysis
in dynare 4.6.3).
I said this is the “problem” (in quotes) because I do not know if it is a bug or it is a feature that I did not understand. Would you help me here?
In addition, when I suppress the command identification
in mod file, the M_.params shows the posterior_mean and generate different IRFs between the two samples.
You are not supposed to check identification
after estimation
. What you report is expected behavior. As documented in the manual, the default parameter_set
for identification
is the prior_mean
. You would need
identification(parameter_set=posterior_mean)
I see what you mean, professor Pfeifer.
That’s my fault.
Thank you so much.